Speculative Scheduling of Parameter Sweep Applications Using Job Behaviour Descriptions - Cloud, Grid and High Performance Computing: Emerging Applications - page 85

Algorithm 2.

void schedule(Job j, JobDescription d) {
Map<CE, Long> m = new HashMap<CE, Long>();
for (ClusterProfile c: g) {
if (c.canRun(descr)) {
m.put(c, estimate(c, d)); // calculate the est. finish time
}
}
CE c = getOptimalCE(m); // get the optimal CE
executeJob(j, c); // run job j on c
}

3. The scheduler applies the proposed schedul-
ing algorithm, which - using the behaviour
description and the information available on
the current state of the Grid - calculates the
estimated job finishing time for each Grid
component, and schedules the job to the
component where the job would be finished
the earliest.
4. The job is executed on a computer belonging
to the chosen Grid component. The resource
consumption of the job is monitored, and
after the job is terminated, the collected
information is used by the description re-
pository service to update the description
repository with a refined description.
5. The output of the job (and the behaviour
description of the job) is copied to the speci-
fied target node.

Manager. The simplified code-snippet in Algo-
rithm 2 presents the static data feeder algorithm.
The estimated execution time of a job described
by d on cluster c is calculated as follows.

m

∑ l
1

C c d ( , ) = ( , ) * ( , d i C c d i )

i =

The actual state of cluster c is obtained from
the GIS. The estimate ( c , d ) estimated termination
time of the given job on cluster c is the sum of the
estimated job execution time C ( c , d ), the ”length”
(measured by) of the job queue on that cluster
( Q ( c )), and the time necessary for preparing the
input files (before running the job) and delivering
result/output files (after the job is terminated):

estimate ( c , d ) = C ( c , d ) + Q ( c ) +
fileTransferTime ( c , d )

Static Data Feeder Strategy

Please note that before running the job on the
chosen cluster the necessary files are replicated
by the Replica Manager (The DataGrid Project).

The static data feeder strategy ranks each Comput-
ing Element (CE) by estimating the termination
time of the submitted job on the given component.
After the ranking of CEs the scheduler runs the job
on the CE with the highest rank, i.e. the earliest
completion time. The estimated job completion
time depends on the job description and on the
information collected from the GIS and the Replica

Dynamic Data Feeder Strategy

The basic idea behind the dynamic data feeder
strategy is to download relevant parts of the input
files (those parts that the job will presumably
access) and to upload the output of the job to the